Data-centric Profiling Working Group Outbrief Basic Concept - - PowerPoint PPT Presentation

Data-centric Profiling Working Group Outbrief

Basic Concept

• Associating performance data with data objects (arrays), beyond

code contexts (loops, procedures)

– PMU support – data-centric attribution – use of data-centric profiling

2

Data-centric Profiling WG

• Current PMU support

– Intel PEBS, AMD IBS, IBM Mark events to sample memory accesses

• effective address, latency, memory layers

– monitoring loads only is not enough, but also stores/prefetching instructions

• use L1D replacement event (https://software.intel.com/en-us/forums/intel-performance-

bottleneck-analyzer/topic/326007)

– better to monitor evicted cache lines

• Jeff’s paper: http://www.cs.umd.edu/~hollings/papers/ijhpca06.pdf

– LBR: use call stack mode (monitoring calls/returns) to reconstruct the call stack

• 16 frames on average with 32 LBR slots

– Intel PT

• ptwrite (Goldmont), a lightweight printf triggers LBR. Call ptwrite inside malloc can obtain

the call path from LBR

– page fault events, a hardware event (Goldmont)

• possible measure first touch location

– limitation

• no PID or TID. OS Kernel needs to get this information
• PEBS latency_above_threshold may produce biased results

– sample MEM_LOAD/MEM_RETIRED

3

• Handle attribution to data structures

– static — easy to handle from symbol table

• need Dyninst to extract allocation source lines from DWARF

– heap

• high overhead if malloc/free are frequently called

– probably use ptwrite to reduce the overhead

• call stack is important

– merge the objects allocated in the same call path – (David) meaningful allocation site may a few frames above the “malloc”

– stack

• (Xiaozhu) Dyninst supports to extract the information from DWARF

4

• Use of data-centric profiling

– locality optimization

• data layout optimization

– David has some work in helping developers change data layout

• temporal locality

– false sharing

• HITM events for loads

– may miss store-store false sharing

• Intel PTU, toplev, Feather (Xu’s group) identify false sharing

– NUMA optimization

• lightweight pattern analysis across threads

– structure splitting

• identify how different fields of a data structure are accessed

– structslim from Xu’s group: https://dl.acm.org/citation.cfm? id=2854053

5

Challenges

• Stephane: how to do data profiling offline

– collect all raw data online with low overhead – perform data attribution offline – timestamp information

• Michael: automate the fix

– Joseph (UPenn)’s approach of detecting and fixing false sharing – Intel PGO can improve a DB workload by 25% to guide global data reorganization on Itanium

• Stephane: compiler support to annotate each memory access

instruction

– which type accessed – the offset

• Michael: data-centric profiling on small cores

– insights for temporal locality

6